2 ms·

The fact that it is an incredibly hard problem that they are making steady progress in and they basically dragged EVs kicking and screaming into public consciou

by hoseja 7y ago

The fact that it is an incredibly hard problem that they are making steady progress in and they basically dragged EVs kicking and screaming into public consciousness and wide availability is just a minor detail! Now how are my puts doing...

fxtentacle 7y ago
I don't agree. Most of the Tesla's vision processing appears to be AIs trained with census loss, meaning the same stuff that Daimler Benz used in 2004 for their lane assistance systems. I recently gave a "state of the art" talk about optical flow and I found it extremely disappointing that current technology still struggles with recognizing shiny surfaces and fences, despite 16 years of hardware progress. But most AI researchers appear to be more interested in re-inventing the wheel on a GPU instead of doing actual mathematical algorithm research. Tesla is doing marginally better than the average due to their custom hardware, which grants them more than the 10 GB that people using NVIDIA GPUs are limited to. But still, Tesla, Skydio, and NVIDIA appear to be using optical flow AI architectures that are fundamentally incapable of performing well, due to being based on the (mathematically provably wrong) assumption that a scale pyramid works for optical flow detection.
- xrayarx 7y ago
  Seems like you made a point. Could you provide more details? What are the other options and who is using them?
  - fxtentacle 7y ago
    Sure, I'm happy to discuss AI in more detail. The problem is that pretty much all AIs are built upon the PWC-Net architecture from Nvidia. The solution would be to use a different architecture, but that doesn't fit into GPU. Here's the Sintel ranking: http://sintel.is.tue.mpg.de/results http://sintel.is.tue.mpg.de/results And if you look at the AIs, they all repurpose this paper: https://arxiv.org/abs/1709.0237 https://arxiv.org/abs/1709.0237 However, you cannot store one optical flow per pixel, because the solution is not unique. For example, that's why Tesla and Skydio have problems with reflective surfaces. You see both the surface and its reflection move, but the AI can only store 1 result out of those 2 movements. In my opinion, the solution is to use Fourier methods, which were shown to work well in a 2000 dagm paper called minimum values of the structure tensor. But nobody has even attempted to port that to GPU yet or use it as part of AI.
    - bdamm 7y ago
      Well Musk did recently state that Tesla is soon to release time-based object "3d labeling", where indeed the movement of objects through space is part of their labeling. [1] [1]: https://electrek.co/2020/03/04/tesla-car-find-larking-spot-reverse-summon/ https://electrek.co/2020/03/04/tesla-car-find-larking-spot-r...

by hoseja 7y ago

fxtentacle 7y ago

I don't agree. Most of the Tesla's vision processing appears to be AIs trained with census loss, meaning the same stuff that Daimler Benz used in 2004 for their lane assistance systems. I recently gave a "state of the art" talk about optical flow and I found it extremely disappointing that current technology still struggles with recognizing shiny surfaces and fences, despite 16 years of hardware progress. But most AI researchers appear to be more interested in re-inventing the wheel on a GPU instead of doing actual mathematical algorithm research. Tesla is doing marginally better than the average due to their custom hardware, which grants them more than the 10 GB that people using NVIDIA GPUs are limited to. But still, Tesla, Skydio, and NVIDIA appear to be using optical flow AI architectures that are fundamentally incapable of performing well, due to being based on the (mathematically provably wrong) assumption that a scale pyramid works for optical flow detection.

xrayarx 7y ago
Seems like you made a point. Could you provide more details? What are the other options and who is using them?
- fxtentacle 7y ago
  Sure, I'm happy to discuss AI in more detail. The problem is that pretty much all AIs are built upon the PWC-Net architecture from Nvidia. The solution would be to use a different architecture, but that doesn't fit into GPU. Here's the Sintel ranking: http://sintel.is.tue.mpg.de/results http://sintel.is.tue.mpg.de/results And if you look at the AIs, they all repurpose this paper: https://arxiv.org/abs/1709.0237 https://arxiv.org/abs/1709.0237 However, you cannot store one optical flow per pixel, because the solution is not unique. For example, that's why Tesla and Skydio have problems with reflective surfaces. You see both the surface and its reflection move, but the AI can only store 1 result out of those 2 movements. In my opinion, the solution is to use Fourier methods, which were shown to work well in a 2000 dagm paper called minimum values of the structure tensor. But nobody has even attempted to port that to GPU yet or use it as part of AI.
  - bdamm 7y ago
    Well Musk did recently state that Tesla is soon to release time-based object "3d labeling", where indeed the movement of objects through space is part of their labeling. [1] [1]: https://electrek.co/2020/03/04/tesla-car-find-larking-spot-reverse-summon/ https://electrek.co/2020/03/04/tesla-car-find-larking-spot-r...